CorNet: Unsupervised Deep Homography Estimation for Agricultural Aerial Imagery

Abstract

Efficient and accurate estimation of homographies among images is the first step in mosaicking crop fields for phenotyping. The current strategy uses sophisticated vehicles that have excellent telemetry to hover over a grid of waypoints, imaging each one. This approach simplifies homography estimation, but precludes more flexible, adaptive protocols that can collect richer information. It also makes aerial phenotyping impractical for many researchers and farmers. We are developing an alternative strategy that uses consumer-grade vehicles, freely flown over a variety of trajectories, to collect video. We have developed an unsupervised deep learning network that estimates the sequence of planar homography matrices of our corn fields from imagery, without using any metadata to correct estimation errors. The vehicle was freely flown using a variety of trajectories and camera views. Our system, CorNet , performed faster than and with comparable accuracy to the gold standard ASIFT algorithm in many challenging cases.